https://uberduck.ai/ logo
Join Discord
Powered by
# machine-learning
  • a
    Download failed
  • r
    That's a much better (albiet not perfect) graph
  • r
    Set the model that's saved in your drive to 
    Anyone with a link
  • a
    I already solved. But the model is not talking. Saying again, I edited the epoch to 20, because I uploaded only 2 wav audios. Does the epoch vary with the amount of audios?
  • a
    This audio is just making a robotic noise
  • r
    (°_°) sounds like metal scraping. Creepy
  • r
    Yes, a lot. Assuming you're using legacy in the pic of the graph?
  • a
    That was the result of the graph.
  • r
    This graph, to which I was referring. Legacy, no?
  • u
    @Amizade | Pony's voice creator set the epochs to something really high and then train until the val loss hits 0.06 and saves
  • a
    Okay
  • a
    But which number for example? Will it depend on the amount of audios?
  • a
    What's legacy?
  • u
    Nah just set it to something you'll never reach (i do 5000) and stop the cell when you get to that val loss
  • u
    Legacy tacotron, it's what you're training on right now
  • r
    A seperate version of the notebook vs Pipeline. Model performance can vary on each one if you do use the same settings.
  • a
    Just to comment, I do the following: I copy the legacy to my drive via colab and then start creating and configuring the model. I learned from YouTube tutorials
  • n
    glitched voice.
  • j
    Wow Google Colab limits are trash now, and buying their hosted runtime didn't work for the Tacotron 2 Pipeline saying it can't find the GPU anyway, loooosers
  • s
    nice zwf beat?
  • w
    he a wapr
  • z
    straight fire
  • s
    thank you king
  • s
    drake released new song woww
  • w
    YeeeeeeeeEeEYeYeyeyYeyeyeyEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEEeEEEEEEEEEEEEEEEEEEEEEEEEE. Mehazaids (im aware of the mispell dont mention it)
  • j
    top ten rappers eminem is afraid to diss
  • s
    did anybody else training models notice access restrictions on programmatic access to one of the hifigan checkpoints, this line gives me permission errors 
    gdown.download(d+"1O63eHZR9t1haCdRHQcEgMfMNxiOciSru", os.path.join(output_dir, "HiFiGAN", "do_00000000"), quiet=False)
  • j
    Upgrade gdown
  • j
    Copy code
    !pip install git+https://github.com/wkentaro/gdown.git
  • c
    Do you guys know how can I get better training speeds with crap cpu?
1...101510161017...1068Latest